An analysis of the general properties and convergence of a connectionist model employing probability data as weights and parameters.

摘要

Based on the concept of virtual lateral inhibition, a two layered connectionist model is developed, and its properties explored. This model is called LIBRA/RX. The flow of activation in this model is described by a set of 3N ordinary non-linear differential equations, where N is the number of nodes on the nets' upper level. The mathematical properties of the equations are explored, and in particular, the dynamics of the net is demonstrated to be convergent in nearly all cases. This model has thus far been employed in the task of pattern recognition. In this case, the lower level or input nodes represent the possible features, and the upper level or output nodes represent the possible classes of patterns. This model uses the probability data of the patterns given the features and the features given the patterns as weights. The probabilities of occurrence of the features and patterns appear as parameters of the set of equations.;The equations have been programmed using a Runga-Kutta fourth order method of integration, and this program extensively tested in order to ascertain the suitability of the net in problems of an engineering nature. The first set of tests were conducted using a set of the ten digits as patterns, whose features consisted of strokes modeled after the Rumelhart-Siple fonts. As the system proved capable of recognizing these patterns given the fonts, a further test of the system was devised. The performance of the LIBRA/RX system was contrasted to the performance of nets of linear perceptrons, non-linear perceptrons, and also to that of a backpropagation net, in those nets ability to recognize patterns of RADAR signals which had been classified as either good or bad. The performance of the LIBRA/RX system was on a par with the other three systems, exceeding that of the linear and non-linear perceptrons in two of the three categories considered, and that of the backpropagation net in one of the three categories.